A Functional Operator for Model Uncertainty Quantification in the RKHS

We propose a framework for single-shot predictive uncertainty quantification of a neural network that replaces the conventional Bayesian notion of weight probability density function (PDF) with a functional defined on the model weights in a reproducing kernel Hilbert space (RKHS). The resulting RKHS based analysis yields a potential field based interpretation of the model weight PDF, which allows us to use a perturbation theory based approach in the RKHS to formulate a moment decomposition problem over the model weight-output uncertainty. We show that the extracted moments from this approach automatically decompose the weight PDF around the local neighborhood of the specified model output and determine with great sensitivity the local heterogeneity and anisotropy of the weight PDF around a given model prediction output. Consequently, these functional moments provide much sharper estimates of model predictive uncertainty than the central stochastic moments characterized by Bayesian and ensemble methods. We demonstrate this experimentally by evaluating the error detection capability of the model uncertainty quantification methods on test data that has undergone a covariate shift away from the training PDF learned by the model. We find our proposed measure for uncertainty quantification to be significantly more precise and better calibrated than baseline methods on various benchmark datasets, while also being much faster to compute.